Ever felt the need to deploy your new shiny Pytorch model on the Web / mobile devices / microcontrollers, and the native infrastructure just doesn’t cut it? Meanwhile, the Tensorflow team has been putting impressive work in their inference engines lately.
So, after suffering through the standard conversion process (via ONNX) for quite some time, I decided to make a dedicated library containing as little bullshit as possible.
Hence, Nobuco. It’s designed with simplicity and hackability in mind and automatically solves mismatching channel orders offering near-optimal performance. Try it, and spread the word if you like it. If not, feel free to open an issue or feature request. All forms of contribution are highly welcome!